Walter Gerych’s research while affiliated with Massachusetts Institute of Technology and other places

Developments in deep neural nets have trended towards increasingly larger overparameterized architectures, resulting in lengthy training sessions with ever more elusive training dynamics. Thus, ensuring these models learn accurate generalizable representations of data efficiently is challenging. Previous works have developed specialized techniques from data-pruning, architecture selection, pseudo-label generation, bias identification, or label refurbishment to improve downstream training. Problematically, most methods require prohibitively expensive iterative model training. In this paper, we demonstrate that we can exploit the recent neural tangent kernel (NTK) theory for understanding and improving model training behavior before ever training a model. First, we show a powerful signal derived from the NTK theory can be computed remarkably fast. We then leverage this signal for the design of a unified suite of surprisingly effective tools for the four important tasks of architecture selection, pseudo-label verification, bias identification, and label refurbishment, all requiring zero model training.

An interesting behavior in large language models (LLMs) is prompt sensitivity. When provided with different but semantically equivalent versions of the same prompt, models may produce very different distributions of answers. This suggests that the uncertainty reflected in a model's output distribution for one prompt may not reflect the model's uncertainty about the meaning of the prompt. We model prompt sensitivity as a type of generalization error, and show that sampling across the semantic concept space with paraphrasing perturbations improves uncertainty calibration without compromising accuracy. Additionally, we introduce a new metric for uncertainty decomposition in black-box LLMs that improves upon entropy-based decomposition by modeling semantic continuities in natural language generation. We show that this decomposition metric can be used to quantify how much LLM uncertainty is attributed to prompt sensitivity. Our work introduces a new way to improve uncertainty calibration in prompt-sensitive language models, and provides evidence that some LLMs fail to exhibit consistent general reasoning about the meanings of their inputs.

Corrosion of materials impacts critical economic sectors from infrastructure to transportation. The development of safe, corrosion-inhibiting materials is thus an important area of study in materials science. Traditional corrosion science, preparing and monitoring materials under adverse conditions, is labor intensive and extremely costly. While deep learning has become popular in automating engineering tasks, the development of deep models for corrosion assessment is lacking. We study the unique problem of deep domain adaptation (DA) for automated corrosion assessment of corrosion-inhibiting materials. Corrosion data, i.e., photographic images of corroding materials, is abundant when produced in an artificially controlled laboratory, while controlled images from natural outdoor environments are limited. We thus leverage the more readily available artificial-environment indoor data to train a corrosion assessment classifier to transfer it via domain adaptation. In doing so, we can perform well on the smaller, yet more realistic, outdoor corrosion data, without requiring target labels. We empirically evaluate 5 popular DA models on real-world corrosion image data. Further, we design 8 strategies for ensembling these models. Our study finds across evalaution metrics of accuracy, F1-score, and balanced accuracy that ensembled models incorporating both predictions and confidence scores from each DA model outperform individual DA models. They achieve 41% relative improvement in test accuracy compared to a no-DA baseline. Additionally, we perform a failure analysis study of our model to explain its performance.

Spurious features associated with class labels can lead image classifiers to rely on shortcuts that don't generalize well to new domains. This is especially problematic in medical settings, where biased models fail when applied to different hospitals or systems. In such cases, data-driven methods to reduce spurious correlations are preferred, as clinicians can directly validate the modified images. While Denoising Diffusion Probabilistic Models (Diffusion Models) show promise for natural images, they are impractical for medical use due to the difficulty of describing spurious medical features. To address this, we propose Masked Medical Image Inpainting (MaskMedPaint), which uses text-to-image diffusion models to augment training images by inpainting areas outside key classification regions to match the target domain. We demonstrate that MaskMedPaint enhances generalization to target domains across both natural (Waterbirds, iWildCam) and medical (ISIC 2018, Chest X-ray) datasets, given limited unlabeled target images.

Vision-language model (VLM) embeddings have been shown to encode biases present in their training data, such as societal biases that prescribe negative characteristics to members of various racial and gender identities. VLMs are being quickly adopted for a variety of tasks ranging from few-shot classification to text-guided image generation, making debiasing VLM embeddings crucial. Debiasing approaches that fine-tune the VLM often suffer from catastrophic forgetting. On the other hand, fine-tuning-free methods typically utilize a "one-size-fits-all" approach that assumes that correlation with the spurious attribute can be explained using a single linear direction across all possible inputs. In this work, we propose Bend-VLM, a nonlinear, fine-tuning-free approach for VLM embedding debiasing that tailors the debiasing operation to each unique input. This allows for a more flexible debiasing approach. Additionally, we do not require knowledge of the set of inputs a priori to inference time, making our method more appropriate for online, open-set tasks such as retrieval and text guided image generation.

Vision-language models, like CLIP (Contrastive Language Image Pretraining), are becoming increasingly popular for a wide range of multimodal retrieval tasks. However, prior work has shown that large language and deep vision models can learn historical biases contained in their training sets, leading to perpetuation of stereotypes and potential downstream harm. In this work, we conduct a systematic analysis of the social biases that are present in CLIP, with a focus on the interaction between image and text modalities. We first propose a taxonomy of social biases called So-B-IT, which contains 374 words categorized across ten types of bias. Each type can lead to societal harm if associated with a particular demographic group. Using this taxonomy, we examine images retrieved by CLIP from a facial image dataset using each word as part of a prompt. We find that CLIP frequently displays undesirable associations between harmful words and specific demographic groups, such as retrieving mostly pictures of Middle Eastern men when asked to retrieve images of a "terrorist". Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find. Our findings highlight the importance of evaluating and addressing bias in vision-language models, and suggest the need for transparency and fairness-aware curation of large pre-training datasets.

Vision-language models, like CLIP (Contrastive Language Image Pretraining), are becoming increasingly popular for a wide range of multimodal retrieval tasks. However, prior work has shown that large language and deep vision models can learn historical biases contained in their training sets, leading to perpetuation of stereotypes and potential downstream harm. In this work, we conduct a systematic analysis of the social biases that are present in CLIP, with a focus on the interaction between image and text modalities. We first propose a taxonomy of social biases called So-B-It, which contains 374 words categorized across ten types of bias. Each type can lead to societal harm if associated with a particular demographic group. Using this taxonomy, we examine images retrieved by CLIP from a facial image dataset using each word as part of a prompt. We find that CLIP frequently displays undesirable associations between harmful words and specific demographic groups, such as retrieving mostly pictures of Middle Eastern men when asked to retrieve images of a "terrorist". Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find. Our findings highlight the importance of evaluating and addressing bias in vision-language models, and suggest the need for transparency and fairness-aware curation of large pre-training datasets.

Multi-task learning (MTL) is essential for real-world applications that handle multiple tasks simultaneously, such as selfdriving cars. MTL methods improve the performance of all tasks by utilizing information across tasks to learn a robust shared representation. However, acquiring sufficient labeled data tends to be extremely expensive, especially when having to support many tasks. Recently, Knowledge Amalgamation (KA) has emerged as an effective strategy for addressing the lack of labels by instead learning directly from pretrained models (teachers). KA learns one unified multi-task student that masters all tasks across all teachers. Existing KA for MTL works are limited to teachers with identical architectures, and thus propose layer-to-layer based approaches. Unfortunately, in practice, teachers may have heterogeneous architectures; their layers may not be aligned and their dimensionalities or scales may be incompatible. Amalgamating multi-task teachers with heterogeneous architectures remains an open problem. For this, we design Versatile Common Feature Consolidator (VENUS), the first solution to this problem. VENUS fuses knowledge from the shared representations of each teacher into one unified generalized representation for all tasks. Specifically, we design the Feature Consolidator network that leverages an array of teacher-specific trainable adaptors. These adaptors enable the student to learn from multiple teachers, even if they have incompatible learned representations. We demonstrate that VENUS outperforms five alternative methods on numerous benchmark datasets across a broad spectrum of experiments.